I post trained Gemma 4 E2B into a decision model, shaped like the perplexity decoder, and beat every other local decision model on my task (it's also like 20x faster). Blog coming soon (hopefully).

Decision models like Jev handle what SQL rules cannot. Here's how data teams use them venturebeat.com/ai/decision-…

In January 2024, 355 clinicians in the United States used OpenEvidence in a given month. In June 2026 it was 703,265. That comes from a study published this week in JAMA by Vishal Patel and Anupam Jena of Mass General Brigham and Harvard. In two and a half years, 952,008 verified US clinicians, 556,200 of them physicians, sent this AI tool 271 million clinical questions. The average active user went from 18 queries a month to almost 40. The heaviest users are young doctors, residents, nurse practitioners and physician assistants, along with oncologists, internists and family doctors. Radiologists, ophthalmologists and orthopedic surgeons barely touch it. What surprised me most is what they ask. Only 13.9% of physician queries are about diagnosis. A third are about treatment and management, what to do next with the patient in front of them. Among emergency physicians, 88% of the questions fall outside emergency medicine. I have worked in fertility medicine for years, and I have rarely seen anything spread this fast. Medicine usually takes years, sometimes decades, to adopt something new. Here hundreds of thousands of clinicians picked up a free tool on their own, one question at a time. The study counts questions, not whether the answers were right or whether patients did better, and three of its seven authors are tied to the company. That is the study we need next. Doctors already voted. Now medicine has to catch up.

Radiology Holds Three Quarters of Every AI Clearance. Almost None of It Is Being Used. The FDA has authorized 1,451 AI enabled medical devices. 1,104 of them are radiology. 76% of everything ever cleared, concentrated in one specialty. Now the second number. When researchers traced medical AI through US insurance claims, they found 16 billable AI procedures in total, against more than 500 cleared devices at the time. Only two had passed 10,000 claims: coronary artery disease and diabetic retinopathy. One device had recorded a single claim. Cleared is not deployed. The distance between those two numbers is the entire story. Radiology should have been the easy one. Discrete images, algorithmic outputs, digital standards, a specialty built on pattern recognition. If AI adoption were a technology problem, radiology would be finished. It is not finished, which tells you the problem was never technology. There are four gates, not one. Clearance. Validation. Payment. Deployment. Technology gets you through the first. Governance decides the other three. The MASAI trial, published in The Lancet in January 2026, randomized more than 100,000 women to AI-supported mammography screening or standard double reading. The AI arm detected more cancers with no increase in false positives, cut screen-reading workload by 44% , and at follow-up showed fewer large, invasive and aggressive cancers. Read that carefully. It worked because a national screening program redesigned its reading protocol around the tool. Not because somebody bought the tool. That is the lesson for anyone sitting where I sit. When you buy AI you are not buying a product. You are buying an obligation: to decide when the radiologist defers and when they override, who carries liability when the model misses, who owns the data the algorithm learned from, how you retrain people whose expertise you have just partially automated, and how you get paid for a workflow you have just changed. Most hospitals have not answered those questions. So the license renews, the tool sits idle, and the department buys certification courses instead. The cleanest clinical problem in medicine is still hard to deploy at scale. That is not a verdict on the technology. It is a verdict on us. #HealthcareAI #DigitalHealth #Radiology #KingsJeddah

Vega Universal decision model. Now you don't need to train a separate model for each decision; you can use one engine and multiple LLMs as needed. This solves an important problem current decision models face: accuracy and calibration error. The idea is simple: you train a small projection layer and use that to attach any llm, vegaml package automatically ports that to a common dimension. That's it: joint training with invariance loss embedding.in/blog/vega-u-one… github.com/NandhaKishorM/veg… huggingface.co/nandakishorm/…

Do we need decision models to also think in non-holonomic contexts?

Must-read paper integrating decision models like Jev into harnesses. They distilled an open 4B decision model, mostly from Jev's answers on the agent's own states, that beats Jev inside that agent and is much cheaper to run. Their harness, JevAgentHarness, sends each routine choice to the decision model and uses its answer only when the confidence clears a threshold fitted on the agent's own trajectories. Below the threshold, the LLM decides as usual. With the fitted threshold, tool and argument cascades cut cost by 14.2% without changing the outcome of any of 500 tasks. A fixed threshold of 0.5 lost 21.5 points. They also train AgentDecision-4B on states the agent reaches after its own mistakes, with Jev's answers as labels. On tasks unlike its training data, the agent solved 91.5% with it and 82.0% with Jev. On familiar tasks, 4B models trained this way made agents 81.1 to 89.0% cheaper and up to 2.5 times faster while staying within 3 points of success. Paper: academy.dair.ai/papers/jevag…